Merge GH #1213 Improve ARM processor detection for systems without getauxval

1 files changed, 214 insertions, 0 deletions

diff --git a/src/lib/utils/cpuid/cpuid_arm.cpp b/src/lib/utils/cpuid/cpuid_arm.cpp
new file mode 100644
index 000000000..39b6db652
--- /dev/null
+++ b/src/lib/utils/cpuid/cpuid_arm.cpp
@@ -0,0 +1,214 @@
+/*
+* Runtime CPU detection for ARM
+* (C) 2009,2010,2013,2017 Jack Lloyd
+*
+* Botan is released under the Simplified BSD License (see license.txt)
+*/
+
+#include <botan/cpuid.h>
+
+#if defined(BOTAN_TARGET_CPU_IS_ARM_FAMILY)
+
+#if defined(BOTAN_TARGET_OS_HAS_GETAUXVAL)
+  #include <sys/auxv.h>
+
+#elif defined(BOTAN_TARGET_OS_IS_IOS)
+  #include <sys/types.h>
+  #include <sys/sysctl.h>
+
+#else
+  #include <botan/internal/os_utils.h>
+
+#endif
+
+#endif
+
+namespace Botan {
+
+#if defined(BOTAN_TARGET_CPU_IS_ARM_FAMILY)
+
+#if defined(BOTAN_TARGET_OS_IS_IOS)
+
+namespace {
+
+uint64_t flags_by_ios_machine_type(const std::string& machine)
+   {
+   /*
+   * This relies on a map of known machine names to features. This
+   * will quickly grow out of date as new products are introduced, but
+   * is apparently the best we can do for iOS.
+   */
+
+   struct version_info {
+      std::string name;
+      size_t min_version_neon;
+      size_t min_version_armv8;
+      };
+
+   static const version_info min_versions[] = {
+      { "iPhone", 2, 6 },
+      { "iPad", 1, 4 },
+      { "iPod", 4, 7 },
+      { "AppleTV", 2, 5 },
+   };
+
+   if(machine.size() < 3)
+      return 0;
+
+   auto comma = machine.find(',');
+
+   // Simulator, or something we don't know about
+   if(comma == std::string::npos)
+      return 0;
+
+   std::string product = machine.substr(0, comma);
+
+   size_t version = 0;
+   size_t place = 1;
+   while(product.size() > 1 && ::isdigit(product.back()))
+      {
+      const size_t digit = product.back() - '0';
+      version += digit * place;
+      place *= 10;
+      product.pop_back();
+      }
+
+   if(version == 0)
+      return 0;
+
+   for(const version_info& info : min_versions)
+      {
+      if(info.name != product)
+         continue;
+
+      if(version >= info.min_version_armv8)
+         {
+         return CPUID::CPUID_ARM_AES_BIT |
+                CPUID::CPUID_ARM_PMULL_BIT |
+                CPUID::CPUID_ARM_SHA1_BIT |
+                CPUID::CPUID_ARM_SHA2_BIT |
+                CPUID::CPUID_ARM_NEON_BIT;
+         }
+
+      if(version >= info.min_version_neon)
+         return CPUID::CPUID_ARM_NEON_BIT;
+      }
+
+   // Some other product we don't know about
+   return 0;
+   }
+
+}
+
+#endif
+
+uint64_t CPUID::detect_cpu_features(size_t* cache_line_size)
+   {
+   uint64_t detected_features = 0;
+
+#if defined(BOTAN_TARGET_OS_HAS_GETAUXVAL)
+   /*
+   * On systems with getauxval these bits should normally be defined
+   * in bits/auxv.h but some buggy? glibc installs seem to miss them.
+   * These following values are all fixed, for the Linux ELF format,
+   * so we just hardcode them in ARM_hwcap_bit enum.
+   */
+
+   enum ARM_hwcap_bit {
+#if defined(BOTAN_TARGET_ARCH_IS_ARM32)
+      NEON_bit  = (1 << 12),
+      AES_bit   = (1 << 0),
+      PMULL_bit = (1 << 1),
+      SHA1_bit  = (1 << 2),
+      SHA2_bit  = (1 << 3),
+
+      ARCH_hwcap_neon   = 16, // AT_HWCAP
+      ARCH_hwcap_crypto = 26, // AT_HWCAP2
+#elif defined(BOTAN_TARGET_ARCH_IS_ARM64)
+      NEON_bit  = (1 << 1),
+      AES_bit   = (1 << 3),
+      PMULL_bit = (1 << 4),
+      SHA1_bit  = (1 << 5),
+      SHA2_bit  = (1 << 6),
+
+      ARCH_hwcap_neon   = 16, // AT_HWCAP
+      ARCH_hwcap_crypto = 16, // AT_HWCAP
+#endif
+   };
+
+#if defined(AT_DCACHEBSIZE)
+   const unsigned long dcache_line = ::getauxval(AT_DCACHEBSIZE);
+
+   // plausibility check
+   if(dcache_line == 32 || dcache_line == 64 || dcache_line == 128)
+      *cache_line_size = static_cast<size_t>(dcache_line);
+#endif
+
+   const unsigned long hwcap_neon = ::getauxval(ARM_hwcap_bit::ARCH_hwcap_neon);
+   if(hwcap_neon & ARM_hwcap_bit::NEON_bit)
+      detected_features |= CPUID::CPUID_ARM_NEON_BIT;
+
+   /*
+   On aarch64 this ends up calling getauxval twice with AT_HWCAP
+   It doesn't seem worth optimizing this out, since getauxval is
+   just reading a field in the ELF header.
+   */
+   const unsigned long hwcap_crypto = ::getauxval(ARM_hwcap_bit::ARCH_hwcap_crypto);
+   if(hwcap_crypto & ARM_hwcap_bit::AES_bit)
+      detected_features |= CPUID::CPUID_ARM_AES_BIT;
+   if(hwcap_crypto & ARM_hwcap_bit::PMULL_bit)
+      detected_features |= CPUID::CPUID_ARM_PMULL_BIT;
+   if(hwcap_crypto & ARM_hwcap_bit::SHA1_bit)
+      detected_features |= CPUID::CPUID_ARM_SHA1_BIT;
+   if(hwcap_crypto & ARM_hwcap_bit::SHA2_bit)
+      detected_features |= CPUID::CPUID_ARM_SHA2_BIT;
+
+#elif defined(BOTAN_TARGET_OS_IS_IOS)
+
+   char machine[64] = { 0 };
+   size_t size = sizeof(machine) - 1;
+   ::sysctlbyname("hw.machine", machine, &size, nullptr, 0);
+
+   detected_features = flags_by_ios_machine_type(machine);
+
+#elif defined(BOTAN_USE_GCC_INLINE_ASM) && defined(BOTAN_TARGET_ARCH_IS_ARM64)
+
+   /*
+   No getauxval API available, fall back on probe functions. We only
+   bother with Aarch64 here to simplify the code and because going to
+   extreme contortions to support detect NEON on devices that probably
+   don't support it doesn't seem worthwhile.
+
+   NEON registers v0-v7 are caller saved in Aarch64
+   */
+
+   auto neon_probe  = []() -> int { asm("and v0.16b, v0.16b, v0.16b"); return 1; };
+   auto aes_probe   = []() -> int { asm(".word 0x4e284800"); return 1; };
+   auto pmull_probe = []() -> int { asm(".word 0x0ee0e000"); return 1; };
+   auto sha1_probe  = []() -> int { asm(".word 0x5e280800"); return 1; };
+   auto sha2_probe  = []() -> int { asm(".word 0x5e282800"); return 1; };
+
+   // Only bother running the crypto detection if we found NEON
+
+   if(OS::run_cpu_instruction_probe(neon_probe) == 1)
+      {
+      detected_features |= CPUID::CPUID_ARM_NEON_BIT;
+
+      if(OS::run_cpu_instruction_probe(aes_probe) == 1)
+         detected_features |= CPUID::CPUID_ARM_AES_BIT;
+      if(OS::run_cpu_instruction_probe(pmull_probe) == 1)
+         detected_features |= CPUID::CPUID_ARM_PMULL_BIT;
+      if(OS::run_cpu_instruction_probe(sha1_probe) == 1)
+         detected_features |= CPUID::CPUID_ARM_SHA1_BIT;
+      if(OS::run_cpu_instruction_probe(sha2_probe) == 1)
+         detected_features |= CPUID::CPUID_ARM_SHA2_BIT;
+      }
+
+#endif
+
+   return detected_features;
+   }
+
+#endif
+
+}